예제 #1
0
파일: parallel.cpp 프로젝트: dieface/caffe
void P2PSync<Dtype>::run(const vector<device*>& gpus) {
  // Pair devices for map-reduce synchronization
  vector<DevicePair> pairs;
  DevicePair::compute(gpus, &pairs);
  ostringstream s;
  for (int i = 1; i < pairs.size(); ++i) {
    s << (i == 1 ? "" : ", ") << pairs[i].get_parent() << ":"
      << pairs[i].get_device();
  }
  LOG(INFO)<< "GPUs pairs " << s.str();

  SolverParameter param(solver_->param());
  vector<shared_ptr<P2PSync<Dtype> > > syncs(gpus.size());

  // Build the GPU tree by finding the parent for each solver
  for (int attempts = 0; attempts < pairs.size(); ++attempts) {
    for (int i = 1; i < pairs.size(); ++i) {
      if (!syncs[i].get()) {
        P2PSync<Dtype>* parent = NULL;
        for (int j = 0; j < syncs.size(); ++j) {
          P2PSync<Dtype>* sync = j == 0 ? this : syncs[j].get();
          if (sync) {
            const SolverParameter& p = sync->solver()->param();
            if (p.device_id() == pairs[i].get_parent()->list_id()) {
              parent = sync;
            }
          }
        }
        if (parent) {
          param.set_device_id(pairs[i].get_device()->list_id());
          syncs[i].reset(new P2PSync<Dtype>(solver_, parent, param));
          parent->children_.push_back((P2PSync<Dtype>*) syncs[i].get());
        }
      }
    }
  }

  LOG(INFO)<< "Starting Optimization";

  for (int i = 1; i < syncs.size(); ++i) {
    syncs[i]->StartInternalThread(Caffe::GetDefaultDevice());
  }

  // Run root solver on current thread
  solver_->Solve();

  for (int i = 1; i < syncs.size(); ++i) {
    syncs[i]->StopInternalThread();
  }
}
예제 #2
0
파일: parallel.cpp 프로젝트: fireae/caffe
void P2PSync<Dtype>::Prepare(const vector<int>& gpus,
            vector<shared_ptr<P2PSync<Dtype> > >* syncs) {
  // Pair devices for map-reduce synchronization
  vector<DevicePair> pairs;
  DevicePair::compute(gpus, &pairs);
  ostringstream s;
  for (int i = 1; i < pairs.size(); ++i) {
    s << (i == 1 ? "" : ", ") << pairs[i].parent() << ":" << pairs[i].device();
  }
  LOG(INFO)<< "GPUs pairs " << s.str();

  SolverParameter param(solver_->param());

  // Build the GPU tree by finding the parent for each solver
  for (int attempts = 0; attempts < pairs.size(); ++attempts) {
    for (int i = 1; i < pairs.size(); ++i) {
      if (!syncs->at(i).get()) {
        P2PSync<Dtype>* parent = NULL;
        for (int j = 0; j < syncs->size(); ++j) {
          P2PSync<Dtype>* sync = j == 0 ? this : syncs->at(j).get();
          if (sync) {
            const SolverParameter& p = sync->solver()->param();
            if (p.device_id() == pairs[i].parent()) {
              parent = sync;
            }
          }
        }
        if (parent) {
          param.set_device_id(pairs[i].device());
          syncs->at(i).reset(new P2PSync<Dtype>(solver_, parent, param));
          parent->children_.push_back((P2PSync<Dtype>*) syncs->at(i).get());
        }
      }
    }
  }
}